xref: /openbmc/linux/fs/pipe.c (revision 6342a20e)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  *  linux/fs/pipe.c
4  *
5  *  Copyright (C) 1991, 1992, 1999  Linus Torvalds
6  */
7 
8 #include <linux/mm.h>
9 #include <linux/file.h>
10 #include <linux/poll.h>
11 #include <linux/slab.h>
12 #include <linux/module.h>
13 #include <linux/init.h>
14 #include <linux/fs.h>
15 #include <linux/log2.h>
16 #include <linux/mount.h>
17 #include <linux/pseudo_fs.h>
18 #include <linux/magic.h>
19 #include <linux/pipe_fs_i.h>
20 #include <linux/uio.h>
21 #include <linux/highmem.h>
22 #include <linux/pagemap.h>
23 #include <linux/audit.h>
24 #include <linux/syscalls.h>
25 #include <linux/fcntl.h>
26 #include <linux/memcontrol.h>
27 #include <linux/watch_queue.h>
28 #include <linux/sysctl.h>
29 
30 #include <linux/uaccess.h>
31 #include <asm/ioctls.h>
32 
33 #include "internal.h"
34 
35 /*
36  * New pipe buffers will be restricted to this size while the user is exceeding
37  * their pipe buffer quota. The general pipe use case needs at least two
38  * buffers: one for data yet to be read, and one for new data. If this is less
39  * than two, then a write to a non-empty pipe may block even if the pipe is not
40  * full. This can occur with GNU make jobserver or similar uses of pipes as
41  * semaphores: multiple processes may be waiting to write tokens back to the
42  * pipe before reading tokens: https://lore.kernel.org/lkml/1628086770.5rn8p04n6j.none@localhost/.
43  *
44  * Users can reduce their pipe buffers with F_SETPIPE_SZ below this at their
45  * own risk, namely: pipe writes to non-full pipes may block until the pipe is
46  * emptied.
47  */
48 #define PIPE_MIN_DEF_BUFFERS 2
49 
50 /*
51  * The max size that a non-root user is allowed to grow the pipe. Can
52  * be set by root in /proc/sys/fs/pipe-max-size
53  */
54 static unsigned int pipe_max_size = 1048576;
55 
56 /* Maximum allocatable pages per user. Hard limit is unset by default, soft
57  * matches default values.
58  */
59 static unsigned long pipe_user_pages_hard;
60 static unsigned long pipe_user_pages_soft = PIPE_DEF_BUFFERS * INR_OPEN_CUR;
61 
62 /*
63  * We use head and tail indices that aren't masked off, except at the point of
64  * dereference, but rather they're allowed to wrap naturally.  This means there
65  * isn't a dead spot in the buffer, but the ring has to be a power of two and
66  * <= 2^31.
67  * -- David Howells 2019-09-23.
68  *
69  * Reads with count = 0 should always return 0.
70  * -- Julian Bradfield 1999-06-07.
71  *
72  * FIFOs and Pipes now generate SIGIO for both readers and writers.
73  * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
74  *
75  * pipe_read & write cleanup
76  * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
77  */
78 
79 static void pipe_lock_nested(struct pipe_inode_info *pipe, int subclass)
80 {
81 	if (pipe->files)
82 		mutex_lock_nested(&pipe->mutex, subclass);
83 }
84 
85 void pipe_lock(struct pipe_inode_info *pipe)
86 {
87 	/*
88 	 * pipe_lock() nests non-pipe inode locks (for writing to a file)
89 	 */
90 	pipe_lock_nested(pipe, I_MUTEX_PARENT);
91 }
92 EXPORT_SYMBOL(pipe_lock);
93 
94 void pipe_unlock(struct pipe_inode_info *pipe)
95 {
96 	if (pipe->files)
97 		mutex_unlock(&pipe->mutex);
98 }
99 EXPORT_SYMBOL(pipe_unlock);
100 
101 static inline void __pipe_lock(struct pipe_inode_info *pipe)
102 {
103 	mutex_lock_nested(&pipe->mutex, I_MUTEX_PARENT);
104 }
105 
106 static inline void __pipe_unlock(struct pipe_inode_info *pipe)
107 {
108 	mutex_unlock(&pipe->mutex);
109 }
110 
111 void pipe_double_lock(struct pipe_inode_info *pipe1,
112 		      struct pipe_inode_info *pipe2)
113 {
114 	BUG_ON(pipe1 == pipe2);
115 
116 	if (pipe1 < pipe2) {
117 		pipe_lock_nested(pipe1, I_MUTEX_PARENT);
118 		pipe_lock_nested(pipe2, I_MUTEX_CHILD);
119 	} else {
120 		pipe_lock_nested(pipe2, I_MUTEX_PARENT);
121 		pipe_lock_nested(pipe1, I_MUTEX_CHILD);
122 	}
123 }
124 
125 static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
126 				  struct pipe_buffer *buf)
127 {
128 	struct page *page = buf->page;
129 
130 	/*
131 	 * If nobody else uses this page, and we don't already have a
132 	 * temporary page, let's keep track of it as a one-deep
133 	 * allocation cache. (Otherwise just release our reference to it)
134 	 */
135 	if (page_count(page) == 1 && !pipe->tmp_page)
136 		pipe->tmp_page = page;
137 	else
138 		put_page(page);
139 }
140 
141 static bool anon_pipe_buf_try_steal(struct pipe_inode_info *pipe,
142 		struct pipe_buffer *buf)
143 {
144 	struct page *page = buf->page;
145 
146 	if (page_count(page) != 1)
147 		return false;
148 	memcg_kmem_uncharge_page(page, 0);
149 	__SetPageLocked(page);
150 	return true;
151 }
152 
153 /**
154  * generic_pipe_buf_try_steal - attempt to take ownership of a &pipe_buffer
155  * @pipe:	the pipe that the buffer belongs to
156  * @buf:	the buffer to attempt to steal
157  *
158  * Description:
159  *	This function attempts to steal the &struct page attached to
160  *	@buf. If successful, this function returns 0 and returns with
161  *	the page locked. The caller may then reuse the page for whatever
162  *	he wishes; the typical use is insertion into a different file
163  *	page cache.
164  */
165 bool generic_pipe_buf_try_steal(struct pipe_inode_info *pipe,
166 		struct pipe_buffer *buf)
167 {
168 	struct page *page = buf->page;
169 
170 	/*
171 	 * A reference of one is golden, that means that the owner of this
172 	 * page is the only one holding a reference to it. lock the page
173 	 * and return OK.
174 	 */
175 	if (page_count(page) == 1) {
176 		lock_page(page);
177 		return true;
178 	}
179 	return false;
180 }
181 EXPORT_SYMBOL(generic_pipe_buf_try_steal);
182 
183 /**
184  * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
185  * @pipe:	the pipe that the buffer belongs to
186  * @buf:	the buffer to get a reference to
187  *
188  * Description:
189  *	This function grabs an extra reference to @buf. It's used in
190  *	the tee() system call, when we duplicate the buffers in one
191  *	pipe into another.
192  */
193 bool generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
194 {
195 	return try_get_page(buf->page);
196 }
197 EXPORT_SYMBOL(generic_pipe_buf_get);
198 
199 /**
200  * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
201  * @pipe:	the pipe that the buffer belongs to
202  * @buf:	the buffer to put a reference to
203  *
204  * Description:
205  *	This function releases a reference to @buf.
206  */
207 void generic_pipe_buf_release(struct pipe_inode_info *pipe,
208 			      struct pipe_buffer *buf)
209 {
210 	put_page(buf->page);
211 }
212 EXPORT_SYMBOL(generic_pipe_buf_release);
213 
214 static const struct pipe_buf_operations anon_pipe_buf_ops = {
215 	.release	= anon_pipe_buf_release,
216 	.try_steal	= anon_pipe_buf_try_steal,
217 	.get		= generic_pipe_buf_get,
218 };
219 
220 /* Done while waiting without holding the pipe lock - thus the READ_ONCE() */
221 static inline bool pipe_readable(const struct pipe_inode_info *pipe)
222 {
223 	unsigned int head = READ_ONCE(pipe->head);
224 	unsigned int tail = READ_ONCE(pipe->tail);
225 	unsigned int writers = READ_ONCE(pipe->writers);
226 
227 	return !pipe_empty(head, tail) || !writers;
228 }
229 
230 static ssize_t
231 pipe_read(struct kiocb *iocb, struct iov_iter *to)
232 {
233 	size_t total_len = iov_iter_count(to);
234 	struct file *filp = iocb->ki_filp;
235 	struct pipe_inode_info *pipe = filp->private_data;
236 	bool was_full, wake_next_reader = false;
237 	ssize_t ret;
238 
239 	/* Null read succeeds. */
240 	if (unlikely(total_len == 0))
241 		return 0;
242 
243 	ret = 0;
244 	__pipe_lock(pipe);
245 
246 	/*
247 	 * We only wake up writers if the pipe was full when we started
248 	 * reading in order to avoid unnecessary wakeups.
249 	 *
250 	 * But when we do wake up writers, we do so using a sync wakeup
251 	 * (WF_SYNC), because we want them to get going and generate more
252 	 * data for us.
253 	 */
254 	was_full = pipe_full(pipe->head, pipe->tail, pipe->max_usage);
255 	for (;;) {
256 		/* Read ->head with a barrier vs post_one_notification() */
257 		unsigned int head = smp_load_acquire(&pipe->head);
258 		unsigned int tail = pipe->tail;
259 		unsigned int mask = pipe->ring_size - 1;
260 
261 #ifdef CONFIG_WATCH_QUEUE
262 		if (pipe->note_loss) {
263 			struct watch_notification n;
264 
265 			if (total_len < 8) {
266 				if (ret == 0)
267 					ret = -ENOBUFS;
268 				break;
269 			}
270 
271 			n.type = WATCH_TYPE_META;
272 			n.subtype = WATCH_META_LOSS_NOTIFICATION;
273 			n.info = watch_sizeof(n);
274 			if (copy_to_iter(&n, sizeof(n), to) != sizeof(n)) {
275 				if (ret == 0)
276 					ret = -EFAULT;
277 				break;
278 			}
279 			ret += sizeof(n);
280 			total_len -= sizeof(n);
281 			pipe->note_loss = false;
282 		}
283 #endif
284 
285 		if (!pipe_empty(head, tail)) {
286 			struct pipe_buffer *buf = &pipe->bufs[tail & mask];
287 			size_t chars = buf->len;
288 			size_t written;
289 			int error;
290 
291 			if (chars > total_len) {
292 				if (buf->flags & PIPE_BUF_FLAG_WHOLE) {
293 					if (ret == 0)
294 						ret = -ENOBUFS;
295 					break;
296 				}
297 				chars = total_len;
298 			}
299 
300 			error = pipe_buf_confirm(pipe, buf);
301 			if (error) {
302 				if (!ret)
303 					ret = error;
304 				break;
305 			}
306 
307 			written = copy_page_to_iter(buf->page, buf->offset, chars, to);
308 			if (unlikely(written < chars)) {
309 				if (!ret)
310 					ret = -EFAULT;
311 				break;
312 			}
313 			ret += chars;
314 			buf->offset += chars;
315 			buf->len -= chars;
316 
317 			/* Was it a packet buffer? Clean up and exit */
318 			if (buf->flags & PIPE_BUF_FLAG_PACKET) {
319 				total_len = chars;
320 				buf->len = 0;
321 			}
322 
323 			if (!buf->len) {
324 				pipe_buf_release(pipe, buf);
325 				spin_lock_irq(&pipe->rd_wait.lock);
326 #ifdef CONFIG_WATCH_QUEUE
327 				if (buf->flags & PIPE_BUF_FLAG_LOSS)
328 					pipe->note_loss = true;
329 #endif
330 				tail++;
331 				pipe->tail = tail;
332 				spin_unlock_irq(&pipe->rd_wait.lock);
333 			}
334 			total_len -= chars;
335 			if (!total_len)
336 				break;	/* common path: read succeeded */
337 			if (!pipe_empty(head, tail))	/* More to do? */
338 				continue;
339 		}
340 
341 		if (!pipe->writers)
342 			break;
343 		if (ret)
344 			break;
345 		if ((filp->f_flags & O_NONBLOCK) ||
346 		    (iocb->ki_flags & IOCB_NOWAIT)) {
347 			ret = -EAGAIN;
348 			break;
349 		}
350 		__pipe_unlock(pipe);
351 
352 		/*
353 		 * We only get here if we didn't actually read anything.
354 		 *
355 		 * However, we could have seen (and removed) a zero-sized
356 		 * pipe buffer, and might have made space in the buffers
357 		 * that way.
358 		 *
359 		 * You can't make zero-sized pipe buffers by doing an empty
360 		 * write (not even in packet mode), but they can happen if
361 		 * the writer gets an EFAULT when trying to fill a buffer
362 		 * that already got allocated and inserted in the buffer
363 		 * array.
364 		 *
365 		 * So we still need to wake up any pending writers in the
366 		 * _very_ unlikely case that the pipe was full, but we got
367 		 * no data.
368 		 */
369 		if (unlikely(was_full))
370 			wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
371 		kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
372 
373 		/*
374 		 * But because we didn't read anything, at this point we can
375 		 * just return directly with -ERESTARTSYS if we're interrupted,
376 		 * since we've done any required wakeups and there's no need
377 		 * to mark anything accessed. And we've dropped the lock.
378 		 */
379 		if (wait_event_interruptible_exclusive(pipe->rd_wait, pipe_readable(pipe)) < 0)
380 			return -ERESTARTSYS;
381 
382 		__pipe_lock(pipe);
383 		was_full = pipe_full(pipe->head, pipe->tail, pipe->max_usage);
384 		wake_next_reader = true;
385 	}
386 	if (pipe_empty(pipe->head, pipe->tail))
387 		wake_next_reader = false;
388 	__pipe_unlock(pipe);
389 
390 	if (was_full)
391 		wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
392 	if (wake_next_reader)
393 		wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
394 	kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
395 	if (ret > 0)
396 		file_accessed(filp);
397 	return ret;
398 }
399 
400 static inline int is_packetized(struct file *file)
401 {
402 	return (file->f_flags & O_DIRECT) != 0;
403 }
404 
405 /* Done while waiting without holding the pipe lock - thus the READ_ONCE() */
406 static inline bool pipe_writable(const struct pipe_inode_info *pipe)
407 {
408 	unsigned int head = READ_ONCE(pipe->head);
409 	unsigned int tail = READ_ONCE(pipe->tail);
410 	unsigned int max_usage = READ_ONCE(pipe->max_usage);
411 
412 	return !pipe_full(head, tail, max_usage) ||
413 		!READ_ONCE(pipe->readers);
414 }
415 
416 static ssize_t
417 pipe_write(struct kiocb *iocb, struct iov_iter *from)
418 {
419 	struct file *filp = iocb->ki_filp;
420 	struct pipe_inode_info *pipe = filp->private_data;
421 	unsigned int head;
422 	ssize_t ret = 0;
423 	size_t total_len = iov_iter_count(from);
424 	ssize_t chars;
425 	bool was_empty = false;
426 	bool wake_next_writer = false;
427 
428 	/* Null write succeeds. */
429 	if (unlikely(total_len == 0))
430 		return 0;
431 
432 	__pipe_lock(pipe);
433 
434 	if (!pipe->readers) {
435 		send_sig(SIGPIPE, current, 0);
436 		ret = -EPIPE;
437 		goto out;
438 	}
439 
440 #ifdef CONFIG_WATCH_QUEUE
441 	if (pipe->watch_queue) {
442 		ret = -EXDEV;
443 		goto out;
444 	}
445 #endif
446 
447 	/*
448 	 * If it wasn't empty we try to merge new data into
449 	 * the last buffer.
450 	 *
451 	 * That naturally merges small writes, but it also
452 	 * page-aligns the rest of the writes for large writes
453 	 * spanning multiple pages.
454 	 */
455 	head = pipe->head;
456 	was_empty = pipe_empty(head, pipe->tail);
457 	chars = total_len & (PAGE_SIZE-1);
458 	if (chars && !was_empty) {
459 		unsigned int mask = pipe->ring_size - 1;
460 		struct pipe_buffer *buf = &pipe->bufs[(head - 1) & mask];
461 		int offset = buf->offset + buf->len;
462 
463 		if ((buf->flags & PIPE_BUF_FLAG_CAN_MERGE) &&
464 		    offset + chars <= PAGE_SIZE) {
465 			ret = pipe_buf_confirm(pipe, buf);
466 			if (ret)
467 				goto out;
468 
469 			ret = copy_page_from_iter(buf->page, offset, chars, from);
470 			if (unlikely(ret < chars)) {
471 				ret = -EFAULT;
472 				goto out;
473 			}
474 
475 			buf->len += ret;
476 			if (!iov_iter_count(from))
477 				goto out;
478 		}
479 	}
480 
481 	for (;;) {
482 		if (!pipe->readers) {
483 			send_sig(SIGPIPE, current, 0);
484 			if (!ret)
485 				ret = -EPIPE;
486 			break;
487 		}
488 
489 		head = pipe->head;
490 		if (!pipe_full(head, pipe->tail, pipe->max_usage)) {
491 			unsigned int mask = pipe->ring_size - 1;
492 			struct pipe_buffer *buf = &pipe->bufs[head & mask];
493 			struct page *page = pipe->tmp_page;
494 			int copied;
495 
496 			if (!page) {
497 				page = alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
498 				if (unlikely(!page)) {
499 					ret = ret ? : -ENOMEM;
500 					break;
501 				}
502 				pipe->tmp_page = page;
503 			}
504 
505 			/* Allocate a slot in the ring in advance and attach an
506 			 * empty buffer.  If we fault or otherwise fail to use
507 			 * it, either the reader will consume it or it'll still
508 			 * be there for the next write.
509 			 */
510 			spin_lock_irq(&pipe->rd_wait.lock);
511 
512 			head = pipe->head;
513 			if (pipe_full(head, pipe->tail, pipe->max_usage)) {
514 				spin_unlock_irq(&pipe->rd_wait.lock);
515 				continue;
516 			}
517 
518 			pipe->head = head + 1;
519 			spin_unlock_irq(&pipe->rd_wait.lock);
520 
521 			/* Insert it into the buffer array */
522 			buf = &pipe->bufs[head & mask];
523 			buf->page = page;
524 			buf->ops = &anon_pipe_buf_ops;
525 			buf->offset = 0;
526 			buf->len = 0;
527 			if (is_packetized(filp))
528 				buf->flags = PIPE_BUF_FLAG_PACKET;
529 			else
530 				buf->flags = PIPE_BUF_FLAG_CAN_MERGE;
531 			pipe->tmp_page = NULL;
532 
533 			copied = copy_page_from_iter(page, 0, PAGE_SIZE, from);
534 			if (unlikely(copied < PAGE_SIZE && iov_iter_count(from))) {
535 				if (!ret)
536 					ret = -EFAULT;
537 				break;
538 			}
539 			ret += copied;
540 			buf->offset = 0;
541 			buf->len = copied;
542 
543 			if (!iov_iter_count(from))
544 				break;
545 		}
546 
547 		if (!pipe_full(head, pipe->tail, pipe->max_usage))
548 			continue;
549 
550 		/* Wait for buffer space to become available. */
551 		if ((filp->f_flags & O_NONBLOCK) ||
552 		    (iocb->ki_flags & IOCB_NOWAIT)) {
553 			if (!ret)
554 				ret = -EAGAIN;
555 			break;
556 		}
557 		if (signal_pending(current)) {
558 			if (!ret)
559 				ret = -ERESTARTSYS;
560 			break;
561 		}
562 
563 		/*
564 		 * We're going to release the pipe lock and wait for more
565 		 * space. We wake up any readers if necessary, and then
566 		 * after waiting we need to re-check whether the pipe
567 		 * become empty while we dropped the lock.
568 		 */
569 		__pipe_unlock(pipe);
570 		if (was_empty)
571 			wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
572 		kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
573 		wait_event_interruptible_exclusive(pipe->wr_wait, pipe_writable(pipe));
574 		__pipe_lock(pipe);
575 		was_empty = pipe_empty(pipe->head, pipe->tail);
576 		wake_next_writer = true;
577 	}
578 out:
579 	if (pipe_full(pipe->head, pipe->tail, pipe->max_usage))
580 		wake_next_writer = false;
581 	__pipe_unlock(pipe);
582 
583 	/*
584 	 * If we do do a wakeup event, we do a 'sync' wakeup, because we
585 	 * want the reader to start processing things asap, rather than
586 	 * leave the data pending.
587 	 *
588 	 * This is particularly important for small writes, because of
589 	 * how (for example) the GNU make jobserver uses small writes to
590 	 * wake up pending jobs
591 	 *
592 	 * Epoll nonsensically wants a wakeup whether the pipe
593 	 * was already empty or not.
594 	 */
595 	if (was_empty || pipe->poll_usage)
596 		wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
597 	kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
598 	if (wake_next_writer)
599 		wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
600 	if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) {
601 		int err = file_update_time(filp);
602 		if (err)
603 			ret = err;
604 		sb_end_write(file_inode(filp)->i_sb);
605 	}
606 	return ret;
607 }
608 
609 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
610 {
611 	struct pipe_inode_info *pipe = filp->private_data;
612 	unsigned int count, head, tail, mask;
613 
614 	switch (cmd) {
615 	case FIONREAD:
616 		__pipe_lock(pipe);
617 		count = 0;
618 		head = pipe->head;
619 		tail = pipe->tail;
620 		mask = pipe->ring_size - 1;
621 
622 		while (tail != head) {
623 			count += pipe->bufs[tail & mask].len;
624 			tail++;
625 		}
626 		__pipe_unlock(pipe);
627 
628 		return put_user(count, (int __user *)arg);
629 
630 #ifdef CONFIG_WATCH_QUEUE
631 	case IOC_WATCH_QUEUE_SET_SIZE: {
632 		int ret;
633 		__pipe_lock(pipe);
634 		ret = watch_queue_set_size(pipe, arg);
635 		__pipe_unlock(pipe);
636 		return ret;
637 	}
638 
639 	case IOC_WATCH_QUEUE_SET_FILTER:
640 		return watch_queue_set_filter(
641 			pipe, (struct watch_notification_filter __user *)arg);
642 #endif
643 
644 	default:
645 		return -ENOIOCTLCMD;
646 	}
647 }
648 
649 /* No kernel lock held - fine */
650 static __poll_t
651 pipe_poll(struct file *filp, poll_table *wait)
652 {
653 	__poll_t mask;
654 	struct pipe_inode_info *pipe = filp->private_data;
655 	unsigned int head, tail;
656 
657 	/* Epoll has some historical nasty semantics, this enables them */
658 	WRITE_ONCE(pipe->poll_usage, true);
659 
660 	/*
661 	 * Reading pipe state only -- no need for acquiring the semaphore.
662 	 *
663 	 * But because this is racy, the code has to add the
664 	 * entry to the poll table _first_ ..
665 	 */
666 	if (filp->f_mode & FMODE_READ)
667 		poll_wait(filp, &pipe->rd_wait, wait);
668 	if (filp->f_mode & FMODE_WRITE)
669 		poll_wait(filp, &pipe->wr_wait, wait);
670 
671 	/*
672 	 * .. and only then can you do the racy tests. That way,
673 	 * if something changes and you got it wrong, the poll
674 	 * table entry will wake you up and fix it.
675 	 */
676 	head = READ_ONCE(pipe->head);
677 	tail = READ_ONCE(pipe->tail);
678 
679 	mask = 0;
680 	if (filp->f_mode & FMODE_READ) {
681 		if (!pipe_empty(head, tail))
682 			mask |= EPOLLIN | EPOLLRDNORM;
683 		if (!pipe->writers && filp->f_version != pipe->w_counter)
684 			mask |= EPOLLHUP;
685 	}
686 
687 	if (filp->f_mode & FMODE_WRITE) {
688 		if (!pipe_full(head, tail, pipe->max_usage))
689 			mask |= EPOLLOUT | EPOLLWRNORM;
690 		/*
691 		 * Most Unices do not set EPOLLERR for FIFOs but on Linux they
692 		 * behave exactly like pipes for poll().
693 		 */
694 		if (!pipe->readers)
695 			mask |= EPOLLERR;
696 	}
697 
698 	return mask;
699 }
700 
701 static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
702 {
703 	int kill = 0;
704 
705 	spin_lock(&inode->i_lock);
706 	if (!--pipe->files) {
707 		inode->i_pipe = NULL;
708 		kill = 1;
709 	}
710 	spin_unlock(&inode->i_lock);
711 
712 	if (kill)
713 		free_pipe_info(pipe);
714 }
715 
716 static int
717 pipe_release(struct inode *inode, struct file *file)
718 {
719 	struct pipe_inode_info *pipe = file->private_data;
720 
721 	__pipe_lock(pipe);
722 	if (file->f_mode & FMODE_READ)
723 		pipe->readers--;
724 	if (file->f_mode & FMODE_WRITE)
725 		pipe->writers--;
726 
727 	/* Was that the last reader or writer, but not the other side? */
728 	if (!pipe->readers != !pipe->writers) {
729 		wake_up_interruptible_all(&pipe->rd_wait);
730 		wake_up_interruptible_all(&pipe->wr_wait);
731 		kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
732 		kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
733 	}
734 	__pipe_unlock(pipe);
735 
736 	put_pipe_info(inode, pipe);
737 	return 0;
738 }
739 
740 static int
741 pipe_fasync(int fd, struct file *filp, int on)
742 {
743 	struct pipe_inode_info *pipe = filp->private_data;
744 	int retval = 0;
745 
746 	__pipe_lock(pipe);
747 	if (filp->f_mode & FMODE_READ)
748 		retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
749 	if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
750 		retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
751 		if (retval < 0 && (filp->f_mode & FMODE_READ))
752 			/* this can happen only if on == T */
753 			fasync_helper(-1, filp, 0, &pipe->fasync_readers);
754 	}
755 	__pipe_unlock(pipe);
756 	return retval;
757 }
758 
759 unsigned long account_pipe_buffers(struct user_struct *user,
760 				   unsigned long old, unsigned long new)
761 {
762 	return atomic_long_add_return(new - old, &user->pipe_bufs);
763 }
764 
765 bool too_many_pipe_buffers_soft(unsigned long user_bufs)
766 {
767 	unsigned long soft_limit = READ_ONCE(pipe_user_pages_soft);
768 
769 	return soft_limit && user_bufs > soft_limit;
770 }
771 
772 bool too_many_pipe_buffers_hard(unsigned long user_bufs)
773 {
774 	unsigned long hard_limit = READ_ONCE(pipe_user_pages_hard);
775 
776 	return hard_limit && user_bufs > hard_limit;
777 }
778 
779 bool pipe_is_unprivileged_user(void)
780 {
781 	return !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN);
782 }
783 
784 struct pipe_inode_info *alloc_pipe_info(void)
785 {
786 	struct pipe_inode_info *pipe;
787 	unsigned long pipe_bufs = PIPE_DEF_BUFFERS;
788 	struct user_struct *user = get_current_user();
789 	unsigned long user_bufs;
790 	unsigned int max_size = READ_ONCE(pipe_max_size);
791 
792 	pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL_ACCOUNT);
793 	if (pipe == NULL)
794 		goto out_free_uid;
795 
796 	if (pipe_bufs * PAGE_SIZE > max_size && !capable(CAP_SYS_RESOURCE))
797 		pipe_bufs = max_size >> PAGE_SHIFT;
798 
799 	user_bufs = account_pipe_buffers(user, 0, pipe_bufs);
800 
801 	if (too_many_pipe_buffers_soft(user_bufs) && pipe_is_unprivileged_user()) {
802 		user_bufs = account_pipe_buffers(user, pipe_bufs, PIPE_MIN_DEF_BUFFERS);
803 		pipe_bufs = PIPE_MIN_DEF_BUFFERS;
804 	}
805 
806 	if (too_many_pipe_buffers_hard(user_bufs) && pipe_is_unprivileged_user())
807 		goto out_revert_acct;
808 
809 	pipe->bufs = kcalloc(pipe_bufs, sizeof(struct pipe_buffer),
810 			     GFP_KERNEL_ACCOUNT);
811 
812 	if (pipe->bufs) {
813 		init_waitqueue_head(&pipe->rd_wait);
814 		init_waitqueue_head(&pipe->wr_wait);
815 		pipe->r_counter = pipe->w_counter = 1;
816 		pipe->max_usage = pipe_bufs;
817 		pipe->ring_size = pipe_bufs;
818 		pipe->nr_accounted = pipe_bufs;
819 		pipe->user = user;
820 		mutex_init(&pipe->mutex);
821 		return pipe;
822 	}
823 
824 out_revert_acct:
825 	(void) account_pipe_buffers(user, pipe_bufs, 0);
826 	kfree(pipe);
827 out_free_uid:
828 	free_uid(user);
829 	return NULL;
830 }
831 
832 void free_pipe_info(struct pipe_inode_info *pipe)
833 {
834 	unsigned int i;
835 
836 #ifdef CONFIG_WATCH_QUEUE
837 	if (pipe->watch_queue)
838 		watch_queue_clear(pipe->watch_queue);
839 #endif
840 
841 	(void) account_pipe_buffers(pipe->user, pipe->nr_accounted, 0);
842 	free_uid(pipe->user);
843 	for (i = 0; i < pipe->ring_size; i++) {
844 		struct pipe_buffer *buf = pipe->bufs + i;
845 		if (buf->ops)
846 			pipe_buf_release(pipe, buf);
847 	}
848 #ifdef CONFIG_WATCH_QUEUE
849 	if (pipe->watch_queue)
850 		put_watch_queue(pipe->watch_queue);
851 #endif
852 	if (pipe->tmp_page)
853 		__free_page(pipe->tmp_page);
854 	kfree(pipe->bufs);
855 	kfree(pipe);
856 }
857 
858 static struct vfsmount *pipe_mnt __read_mostly;
859 
860 /*
861  * pipefs_dname() is called from d_path().
862  */
863 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
864 {
865 	return dynamic_dname(buffer, buflen, "pipe:[%lu]",
866 				d_inode(dentry)->i_ino);
867 }
868 
869 static const struct dentry_operations pipefs_dentry_operations = {
870 	.d_dname	= pipefs_dname,
871 };
872 
873 static struct inode * get_pipe_inode(void)
874 {
875 	struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
876 	struct pipe_inode_info *pipe;
877 
878 	if (!inode)
879 		goto fail_inode;
880 
881 	inode->i_ino = get_next_ino();
882 
883 	pipe = alloc_pipe_info();
884 	if (!pipe)
885 		goto fail_iput;
886 
887 	inode->i_pipe = pipe;
888 	pipe->files = 2;
889 	pipe->readers = pipe->writers = 1;
890 	inode->i_fop = &pipefifo_fops;
891 
892 	/*
893 	 * Mark the inode dirty from the very beginning,
894 	 * that way it will never be moved to the dirty
895 	 * list because "mark_inode_dirty()" will think
896 	 * that it already _is_ on the dirty list.
897 	 */
898 	inode->i_state = I_DIRTY;
899 	inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
900 	inode->i_uid = current_fsuid();
901 	inode->i_gid = current_fsgid();
902 	inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
903 
904 	return inode;
905 
906 fail_iput:
907 	iput(inode);
908 
909 fail_inode:
910 	return NULL;
911 }
912 
913 int create_pipe_files(struct file **res, int flags)
914 {
915 	struct inode *inode = get_pipe_inode();
916 	struct file *f;
917 	int error;
918 
919 	if (!inode)
920 		return -ENFILE;
921 
922 	if (flags & O_NOTIFICATION_PIPE) {
923 		error = watch_queue_init(inode->i_pipe);
924 		if (error) {
925 			free_pipe_info(inode->i_pipe);
926 			iput(inode);
927 			return error;
928 		}
929 	}
930 
931 	f = alloc_file_pseudo(inode, pipe_mnt, "",
932 				O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT)),
933 				&pipefifo_fops);
934 	if (IS_ERR(f)) {
935 		free_pipe_info(inode->i_pipe);
936 		iput(inode);
937 		return PTR_ERR(f);
938 	}
939 
940 	f->private_data = inode->i_pipe;
941 
942 	res[0] = alloc_file_clone(f, O_RDONLY | (flags & O_NONBLOCK),
943 				  &pipefifo_fops);
944 	if (IS_ERR(res[0])) {
945 		put_pipe_info(inode, inode->i_pipe);
946 		fput(f);
947 		return PTR_ERR(res[0]);
948 	}
949 	res[0]->private_data = inode->i_pipe;
950 	res[1] = f;
951 	stream_open(inode, res[0]);
952 	stream_open(inode, res[1]);
953 	return 0;
954 }
955 
956 static int __do_pipe_flags(int *fd, struct file **files, int flags)
957 {
958 	int error;
959 	int fdw, fdr;
960 
961 	if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT | O_NOTIFICATION_PIPE))
962 		return -EINVAL;
963 
964 	error = create_pipe_files(files, flags);
965 	if (error)
966 		return error;
967 
968 	error = get_unused_fd_flags(flags);
969 	if (error < 0)
970 		goto err_read_pipe;
971 	fdr = error;
972 
973 	error = get_unused_fd_flags(flags);
974 	if (error < 0)
975 		goto err_fdr;
976 	fdw = error;
977 
978 	audit_fd_pair(fdr, fdw);
979 	fd[0] = fdr;
980 	fd[1] = fdw;
981 	/* pipe groks IOCB_NOWAIT */
982 	files[0]->f_mode |= FMODE_NOWAIT;
983 	files[1]->f_mode |= FMODE_NOWAIT;
984 	return 0;
985 
986  err_fdr:
987 	put_unused_fd(fdr);
988  err_read_pipe:
989 	fput(files[0]);
990 	fput(files[1]);
991 	return error;
992 }
993 
994 int do_pipe_flags(int *fd, int flags)
995 {
996 	struct file *files[2];
997 	int error = __do_pipe_flags(fd, files, flags);
998 	if (!error) {
999 		fd_install(fd[0], files[0]);
1000 		fd_install(fd[1], files[1]);
1001 	}
1002 	return error;
1003 }
1004 
1005 /*
1006  * sys_pipe() is the normal C calling standard for creating
1007  * a pipe. It's not the way Unix traditionally does this, though.
1008  */
1009 static int do_pipe2(int __user *fildes, int flags)
1010 {
1011 	struct file *files[2];
1012 	int fd[2];
1013 	int error;
1014 
1015 	error = __do_pipe_flags(fd, files, flags);
1016 	if (!error) {
1017 		if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
1018 			fput(files[0]);
1019 			fput(files[1]);
1020 			put_unused_fd(fd[0]);
1021 			put_unused_fd(fd[1]);
1022 			error = -EFAULT;
1023 		} else {
1024 			fd_install(fd[0], files[0]);
1025 			fd_install(fd[1], files[1]);
1026 		}
1027 	}
1028 	return error;
1029 }
1030 
1031 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
1032 {
1033 	return do_pipe2(fildes, flags);
1034 }
1035 
1036 SYSCALL_DEFINE1(pipe, int __user *, fildes)
1037 {
1038 	return do_pipe2(fildes, 0);
1039 }
1040 
1041 /*
1042  * This is the stupid "wait for pipe to be readable or writable"
1043  * model.
1044  *
1045  * See pipe_read/write() for the proper kind of exclusive wait,
1046  * but that requires that we wake up any other readers/writers
1047  * if we then do not end up reading everything (ie the whole
1048  * "wake_next_reader/writer" logic in pipe_read/write()).
1049  */
1050 void pipe_wait_readable(struct pipe_inode_info *pipe)
1051 {
1052 	pipe_unlock(pipe);
1053 	wait_event_interruptible(pipe->rd_wait, pipe_readable(pipe));
1054 	pipe_lock(pipe);
1055 }
1056 
1057 void pipe_wait_writable(struct pipe_inode_info *pipe)
1058 {
1059 	pipe_unlock(pipe);
1060 	wait_event_interruptible(pipe->wr_wait, pipe_writable(pipe));
1061 	pipe_lock(pipe);
1062 }
1063 
1064 /*
1065  * This depends on both the wait (here) and the wakeup (wake_up_partner)
1066  * holding the pipe lock, so "*cnt" is stable and we know a wakeup cannot
1067  * race with the count check and waitqueue prep.
1068  *
1069  * Normally in order to avoid races, you'd do the prepare_to_wait() first,
1070  * then check the condition you're waiting for, and only then sleep. But
1071  * because of the pipe lock, we can check the condition before being on
1072  * the wait queue.
1073  *
1074  * We use the 'rd_wait' waitqueue for pipe partner waiting.
1075  */
1076 static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
1077 {
1078 	DEFINE_WAIT(rdwait);
1079 	int cur = *cnt;
1080 
1081 	while (cur == *cnt) {
1082 		prepare_to_wait(&pipe->rd_wait, &rdwait, TASK_INTERRUPTIBLE);
1083 		pipe_unlock(pipe);
1084 		schedule();
1085 		finish_wait(&pipe->rd_wait, &rdwait);
1086 		pipe_lock(pipe);
1087 		if (signal_pending(current))
1088 			break;
1089 	}
1090 	return cur == *cnt ? -ERESTARTSYS : 0;
1091 }
1092 
1093 static void wake_up_partner(struct pipe_inode_info *pipe)
1094 {
1095 	wake_up_interruptible_all(&pipe->rd_wait);
1096 }
1097 
1098 static int fifo_open(struct inode *inode, struct file *filp)
1099 {
1100 	struct pipe_inode_info *pipe;
1101 	bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
1102 	int ret;
1103 
1104 	filp->f_version = 0;
1105 
1106 	spin_lock(&inode->i_lock);
1107 	if (inode->i_pipe) {
1108 		pipe = inode->i_pipe;
1109 		pipe->files++;
1110 		spin_unlock(&inode->i_lock);
1111 	} else {
1112 		spin_unlock(&inode->i_lock);
1113 		pipe = alloc_pipe_info();
1114 		if (!pipe)
1115 			return -ENOMEM;
1116 		pipe->files = 1;
1117 		spin_lock(&inode->i_lock);
1118 		if (unlikely(inode->i_pipe)) {
1119 			inode->i_pipe->files++;
1120 			spin_unlock(&inode->i_lock);
1121 			free_pipe_info(pipe);
1122 			pipe = inode->i_pipe;
1123 		} else {
1124 			inode->i_pipe = pipe;
1125 			spin_unlock(&inode->i_lock);
1126 		}
1127 	}
1128 	filp->private_data = pipe;
1129 	/* OK, we have a pipe and it's pinned down */
1130 
1131 	__pipe_lock(pipe);
1132 
1133 	/* We can only do regular read/write on fifos */
1134 	stream_open(inode, filp);
1135 
1136 	switch (filp->f_mode & (FMODE_READ | FMODE_WRITE)) {
1137 	case FMODE_READ:
1138 	/*
1139 	 *  O_RDONLY
1140 	 *  POSIX.1 says that O_NONBLOCK means return with the FIFO
1141 	 *  opened, even when there is no process writing the FIFO.
1142 	 */
1143 		pipe->r_counter++;
1144 		if (pipe->readers++ == 0)
1145 			wake_up_partner(pipe);
1146 
1147 		if (!is_pipe && !pipe->writers) {
1148 			if ((filp->f_flags & O_NONBLOCK)) {
1149 				/* suppress EPOLLHUP until we have
1150 				 * seen a writer */
1151 				filp->f_version = pipe->w_counter;
1152 			} else {
1153 				if (wait_for_partner(pipe, &pipe->w_counter))
1154 					goto err_rd;
1155 			}
1156 		}
1157 		break;
1158 
1159 	case FMODE_WRITE:
1160 	/*
1161 	 *  O_WRONLY
1162 	 *  POSIX.1 says that O_NONBLOCK means return -1 with
1163 	 *  errno=ENXIO when there is no process reading the FIFO.
1164 	 */
1165 		ret = -ENXIO;
1166 		if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
1167 			goto err;
1168 
1169 		pipe->w_counter++;
1170 		if (!pipe->writers++)
1171 			wake_up_partner(pipe);
1172 
1173 		if (!is_pipe && !pipe->readers) {
1174 			if (wait_for_partner(pipe, &pipe->r_counter))
1175 				goto err_wr;
1176 		}
1177 		break;
1178 
1179 	case FMODE_READ | FMODE_WRITE:
1180 	/*
1181 	 *  O_RDWR
1182 	 *  POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
1183 	 *  This implementation will NEVER block on a O_RDWR open, since
1184 	 *  the process can at least talk to itself.
1185 	 */
1186 
1187 		pipe->readers++;
1188 		pipe->writers++;
1189 		pipe->r_counter++;
1190 		pipe->w_counter++;
1191 		if (pipe->readers == 1 || pipe->writers == 1)
1192 			wake_up_partner(pipe);
1193 		break;
1194 
1195 	default:
1196 		ret = -EINVAL;
1197 		goto err;
1198 	}
1199 
1200 	/* Ok! */
1201 	__pipe_unlock(pipe);
1202 	return 0;
1203 
1204 err_rd:
1205 	if (!--pipe->readers)
1206 		wake_up_interruptible(&pipe->wr_wait);
1207 	ret = -ERESTARTSYS;
1208 	goto err;
1209 
1210 err_wr:
1211 	if (!--pipe->writers)
1212 		wake_up_interruptible_all(&pipe->rd_wait);
1213 	ret = -ERESTARTSYS;
1214 	goto err;
1215 
1216 err:
1217 	__pipe_unlock(pipe);
1218 
1219 	put_pipe_info(inode, pipe);
1220 	return ret;
1221 }
1222 
1223 const struct file_operations pipefifo_fops = {
1224 	.open		= fifo_open,
1225 	.llseek		= no_llseek,
1226 	.read_iter	= pipe_read,
1227 	.write_iter	= pipe_write,
1228 	.poll		= pipe_poll,
1229 	.unlocked_ioctl	= pipe_ioctl,
1230 	.release	= pipe_release,
1231 	.fasync		= pipe_fasync,
1232 	.splice_write	= iter_file_splice_write,
1233 };
1234 
1235 /*
1236  * Currently we rely on the pipe array holding a power-of-2 number
1237  * of pages. Returns 0 on error.
1238  */
1239 unsigned int round_pipe_size(unsigned long size)
1240 {
1241 	if (size > (1U << 31))
1242 		return 0;
1243 
1244 	/* Minimum pipe size, as required by POSIX */
1245 	if (size < PAGE_SIZE)
1246 		return PAGE_SIZE;
1247 
1248 	return roundup_pow_of_two(size);
1249 }
1250 
1251 /*
1252  * Resize the pipe ring to a number of slots.
1253  *
1254  * Note the pipe can be reduced in capacity, but only if the current
1255  * occupancy doesn't exceed nr_slots; if it does, EBUSY will be
1256  * returned instead.
1257  */
1258 int pipe_resize_ring(struct pipe_inode_info *pipe, unsigned int nr_slots)
1259 {
1260 	struct pipe_buffer *bufs;
1261 	unsigned int head, tail, mask, n;
1262 
1263 	bufs = kcalloc(nr_slots, sizeof(*bufs),
1264 		       GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
1265 	if (unlikely(!bufs))
1266 		return -ENOMEM;
1267 
1268 	spin_lock_irq(&pipe->rd_wait.lock);
1269 	mask = pipe->ring_size - 1;
1270 	head = pipe->head;
1271 	tail = pipe->tail;
1272 
1273 	n = pipe_occupancy(head, tail);
1274 	if (nr_slots < n) {
1275 		spin_unlock_irq(&pipe->rd_wait.lock);
1276 		kfree(bufs);
1277 		return -EBUSY;
1278 	}
1279 
1280 	/*
1281 	 * The pipe array wraps around, so just start the new one at zero
1282 	 * and adjust the indices.
1283 	 */
1284 	if (n > 0) {
1285 		unsigned int h = head & mask;
1286 		unsigned int t = tail & mask;
1287 		if (h > t) {
1288 			memcpy(bufs, pipe->bufs + t,
1289 			       n * sizeof(struct pipe_buffer));
1290 		} else {
1291 			unsigned int tsize = pipe->ring_size - t;
1292 			if (h > 0)
1293 				memcpy(bufs + tsize, pipe->bufs,
1294 				       h * sizeof(struct pipe_buffer));
1295 			memcpy(bufs, pipe->bufs + t,
1296 			       tsize * sizeof(struct pipe_buffer));
1297 		}
1298 	}
1299 
1300 	head = n;
1301 	tail = 0;
1302 
1303 	kfree(pipe->bufs);
1304 	pipe->bufs = bufs;
1305 	pipe->ring_size = nr_slots;
1306 	if (pipe->max_usage > nr_slots)
1307 		pipe->max_usage = nr_slots;
1308 	pipe->tail = tail;
1309 	pipe->head = head;
1310 
1311 	spin_unlock_irq(&pipe->rd_wait.lock);
1312 
1313 	/* This might have made more room for writers */
1314 	wake_up_interruptible(&pipe->wr_wait);
1315 	return 0;
1316 }
1317 
1318 /*
1319  * Allocate a new array of pipe buffers and copy the info over. Returns the
1320  * pipe size if successful, or return -ERROR on error.
1321  */
1322 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long arg)
1323 {
1324 	unsigned long user_bufs;
1325 	unsigned int nr_slots, size;
1326 	long ret = 0;
1327 
1328 #ifdef CONFIG_WATCH_QUEUE
1329 	if (pipe->watch_queue)
1330 		return -EBUSY;
1331 #endif
1332 
1333 	size = round_pipe_size(arg);
1334 	nr_slots = size >> PAGE_SHIFT;
1335 
1336 	if (!nr_slots)
1337 		return -EINVAL;
1338 
1339 	/*
1340 	 * If trying to increase the pipe capacity, check that an
1341 	 * unprivileged user is not trying to exceed various limits
1342 	 * (soft limit check here, hard limit check just below).
1343 	 * Decreasing the pipe capacity is always permitted, even
1344 	 * if the user is currently over a limit.
1345 	 */
1346 	if (nr_slots > pipe->max_usage &&
1347 			size > pipe_max_size && !capable(CAP_SYS_RESOURCE))
1348 		return -EPERM;
1349 
1350 	user_bufs = account_pipe_buffers(pipe->user, pipe->nr_accounted, nr_slots);
1351 
1352 	if (nr_slots > pipe->max_usage &&
1353 			(too_many_pipe_buffers_hard(user_bufs) ||
1354 			 too_many_pipe_buffers_soft(user_bufs)) &&
1355 			pipe_is_unprivileged_user()) {
1356 		ret = -EPERM;
1357 		goto out_revert_acct;
1358 	}
1359 
1360 	ret = pipe_resize_ring(pipe, nr_slots);
1361 	if (ret < 0)
1362 		goto out_revert_acct;
1363 
1364 	pipe->max_usage = nr_slots;
1365 	pipe->nr_accounted = nr_slots;
1366 	return pipe->max_usage * PAGE_SIZE;
1367 
1368 out_revert_acct:
1369 	(void) account_pipe_buffers(pipe->user, nr_slots, pipe->nr_accounted);
1370 	return ret;
1371 }
1372 
1373 /*
1374  * Note that i_pipe and i_cdev share the same location, so checking ->i_pipe is
1375  * not enough to verify that this is a pipe.
1376  */
1377 struct pipe_inode_info *get_pipe_info(struct file *file, bool for_splice)
1378 {
1379 	struct pipe_inode_info *pipe = file->private_data;
1380 
1381 	if (file->f_op != &pipefifo_fops || !pipe)
1382 		return NULL;
1383 #ifdef CONFIG_WATCH_QUEUE
1384 	if (for_splice && pipe->watch_queue)
1385 		return NULL;
1386 #endif
1387 	return pipe;
1388 }
1389 
1390 long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1391 {
1392 	struct pipe_inode_info *pipe;
1393 	long ret;
1394 
1395 	pipe = get_pipe_info(file, false);
1396 	if (!pipe)
1397 		return -EBADF;
1398 
1399 	__pipe_lock(pipe);
1400 
1401 	switch (cmd) {
1402 	case F_SETPIPE_SZ:
1403 		ret = pipe_set_size(pipe, arg);
1404 		break;
1405 	case F_GETPIPE_SZ:
1406 		ret = pipe->max_usage * PAGE_SIZE;
1407 		break;
1408 	default:
1409 		ret = -EINVAL;
1410 		break;
1411 	}
1412 
1413 	__pipe_unlock(pipe);
1414 	return ret;
1415 }
1416 
1417 static const struct super_operations pipefs_ops = {
1418 	.destroy_inode = free_inode_nonrcu,
1419 	.statfs = simple_statfs,
1420 };
1421 
1422 /*
1423  * pipefs should _never_ be mounted by userland - too much of security hassle,
1424  * no real gain from having the whole whorehouse mounted. So we don't need
1425  * any operations on the root directory. However, we need a non-trivial
1426  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1427  */
1428 
1429 static int pipefs_init_fs_context(struct fs_context *fc)
1430 {
1431 	struct pseudo_fs_context *ctx = init_pseudo(fc, PIPEFS_MAGIC);
1432 	if (!ctx)
1433 		return -ENOMEM;
1434 	ctx->ops = &pipefs_ops;
1435 	ctx->dops = &pipefs_dentry_operations;
1436 	return 0;
1437 }
1438 
1439 static struct file_system_type pipe_fs_type = {
1440 	.name		= "pipefs",
1441 	.init_fs_context = pipefs_init_fs_context,
1442 	.kill_sb	= kill_anon_super,
1443 };
1444 
1445 #ifdef CONFIG_SYSCTL
1446 static int do_proc_dopipe_max_size_conv(unsigned long *lvalp,
1447 					unsigned int *valp,
1448 					int write, void *data)
1449 {
1450 	if (write) {
1451 		unsigned int val;
1452 
1453 		val = round_pipe_size(*lvalp);
1454 		if (val == 0)
1455 			return -EINVAL;
1456 
1457 		*valp = val;
1458 	} else {
1459 		unsigned int val = *valp;
1460 		*lvalp = (unsigned long) val;
1461 	}
1462 
1463 	return 0;
1464 }
1465 
1466 static int proc_dopipe_max_size(struct ctl_table *table, int write,
1467 				void *buffer, size_t *lenp, loff_t *ppos)
1468 {
1469 	return do_proc_douintvec(table, write, buffer, lenp, ppos,
1470 				 do_proc_dopipe_max_size_conv, NULL);
1471 }
1472 
1473 static struct ctl_table fs_pipe_sysctls[] = {
1474 	{
1475 		.procname	= "pipe-max-size",
1476 		.data		= &pipe_max_size,
1477 		.maxlen		= sizeof(pipe_max_size),
1478 		.mode		= 0644,
1479 		.proc_handler	= proc_dopipe_max_size,
1480 	},
1481 	{
1482 		.procname	= "pipe-user-pages-hard",
1483 		.data		= &pipe_user_pages_hard,
1484 		.maxlen		= sizeof(pipe_user_pages_hard),
1485 		.mode		= 0644,
1486 		.proc_handler	= proc_doulongvec_minmax,
1487 	},
1488 	{
1489 		.procname	= "pipe-user-pages-soft",
1490 		.data		= &pipe_user_pages_soft,
1491 		.maxlen		= sizeof(pipe_user_pages_soft),
1492 		.mode		= 0644,
1493 		.proc_handler	= proc_doulongvec_minmax,
1494 	},
1495 	{ }
1496 };
1497 #endif
1498 
1499 static int __init init_pipe_fs(void)
1500 {
1501 	int err = register_filesystem(&pipe_fs_type);
1502 
1503 	if (!err) {
1504 		pipe_mnt = kern_mount(&pipe_fs_type);
1505 		if (IS_ERR(pipe_mnt)) {
1506 			err = PTR_ERR(pipe_mnt);
1507 			unregister_filesystem(&pipe_fs_type);
1508 		}
1509 	}
1510 #ifdef CONFIG_SYSCTL
1511 	register_sysctl_init("fs", fs_pipe_sysctls);
1512 #endif
1513 	return err;
1514 }
1515 
1516 fs_initcall(init_pipe_fs);
1517